Upload
others
View
4
Download
0
Embed Size (px)
Citation preview
Marriage and Career:The Dynamic Decisions of Young Men
Eric D. Gould∗
Department of EconomicsHebrew University
andCEPR
June 30, 2003
Abstract
This paper estimates the returns to career decisions in the marriage market andthe returns to marital choices in the labor market. Theoretically, investmentsin the labor market could affect the chances of receiving a marriage offer, thetype of offer, and the probability of getting divorced. Also, marital status couldaffect ones outcomes in the labor market, most notably the marriage premiumin wages. To untangle this simultaneous decision-making process, I develop adynamic programming model of the joint career and marital decisions of youngmen between the ages of 16 and 39. The results show that labor market decisionsare strongly inßuenced by their returns in the marriage market. If there wereno returns to career choices in the marriage market, men would tend to workless, study less, and choose blue-collar jobs over white-collar jobs. These resultssuggest that the existing literature underestimates the true returns to humancapital investments by ignoring their returns in the marriage market. In addition,the results show that the marriage premium is much lower than traditional OLSestimates, and is virtually non-existent for higher wage men. This result suggeststhat while marriage may make low wage men more serious about their careers,marriage has little effect on high wage men who are already highly motivated.
Keywords: marriage, divorce, return to education, occupational choice
JEL Codes: J12, J24
∗For helpful discussions and comments, I thank Gadi Barlevy, Michael Keane, Charles Manksi, RobSauer, Chris Taber, Yoram Weiss, and seminar participants at Northwestern, Tel Aviv University, andMaastricht University. I especially want to thank Simcha Srebnik for computer support. I also thankthe Maurice Falk Institute for Þnancial support. Any errors are my own responsibility.
0
1 Introduction
This paper estimates the returns to career decisions in the marriage market and the
returns to marital choices in the labor market. To do this, I develop a dynamic pro-
gramming model of the joint career and marital decisions of young men over time, using
panel data from the NLSY. The model traces out the sequential and joint schooling,
work, occupation, and marital decisions of white males between the ages of 16 and 39.
During this time period, young men have to make critical decisions regarding their ca-
reers and their marital status, and these decisions often interact with one another. For
example, it is well known that marital status is correlated with a mans wage, education
level, and occupation. Divorce rates, age at Þrst marriage, and length of marriage are
also highly correlated with wages, education, and occupational status. These relation-
ships may represent a causal connection between marriage and career choices, or they
all may be correlated with unobservable characteristics of the individual which are de-
termining his joint success in the marriage and labor markets. One of the goals of this
paper is to estimate the structural parameters underlying the interaction of career and
marriage decisions, while controlling for the endogenous selection of individuals into
their chosen career and marriage paths.
Theoretically, the direction of causality between career and marital decisions could
be in both directions. Marriage could make men more motivated, and thereby, increase
their wages or human capital investments. Or, a higher wage or advanced schooling
could increase ones prospects in the marriage market, thus increasing the chances of
attracting a suitable partner for marriage. In addition, labor market outcomes after
marriage could affect the stability and duration of a marriage, and also the prospects
for re-marriage. Therefore, it is reasonable to conjecture that labor market outcomes
are affected not only by marital status, but also that marital status is affected by labor
market outcomes. The goal of this paper is to sort out the magnitude of these causal
effects, while controlling for self-selection and unobserved heterogeneity.
The vast literature on human capital has thus far ignored the idea that men may
be motivated to succeed in the labor market as a way of improving their prospects in
1
the marriage market. However, this notion is consistent with an evolutionary process
whereby men who are more successful in providing for their partners are more likely
to procreate. To the extent that men are motivated to increase their schooling and
work harder in the labor market only for the extra beneÞts they receive in the marriage
market, estimates of the returns to schooling and work experience in a typical wage
regression could underestimate the true private returns to such activities. This is the
Þrst paper to try to estimate the extent to which men are motivated to succeed in
the labor market in order to enhance their marriage prospects and/or to reduce the
chances of divorce. After estimating the structural parameters of the joint marriage
and career decision process, I construct the counterfactual career path of individuals
when there is no return to career decisions in the marriage market. The results of this
experiment reveal a strong inßuence of marriage considerations on the career choices of
men if it were not for the returns in the marriage market, men would work much less,
go to school less, and choose the blue-collar sector over the white-collar sector more
often. These results suggest that traditional estimates of the returns to human capital
investments are underestimated by looking simply at their returns on the labor market,
and demonstrate the importance of considering the joint nature of marriage and career
decisions.
In addition, estimates of the structural model produce unbiased estimates of the
marriage premium, which are approximately 10% for lower ability men but virtually
non-existent for higher ability men. In contrast, the OLS estimate of the marriage
premium with the same sample is 18%. Thus, typical OLS estimates of the marriage
premium appear to be severely biased upwards due to a correlation between marital
status and unobserved ability. In addition, the results show that there is heterogeneity
in the marriage premium: about 10% for low wage men, but essentially zero for higher
wage men. One interpretation of these Þndings is that marriage may make low wage
men more motivated or serious about their careers, while marriage has little effect on
men with higher wages because they are already highly motivated.
The estimation of the structural parameters also allows for a number of interest-
ing counterfactual experiments. For example, the estimates show how an increasing
2
variation of transitory shocks can lead to many of the broad labor market and marriage
market patterns exhibited over the last few decades: increasing non-employment, more
schooling, an occupational shift from blue-collar to white-collar, declining wages in both
occupations, delays in marriage, and increasing divorce rates. Many of these patterns
also are shown to result from a decrease in the demand for blue-collar labor and/or
increasing inequality of potential wives.
To implement the model, the analysis follows 2,155 white men from ages 16 to 39
using data from the NLSY79. In each year of the sample, individuals are categorized
into one of four career sectors: schooling, white-collar, blue-collar, and home. The
latter category includes all workers whose main activity during the school year was not
going to school or working. In each year, an individuals marital status is recorded
as one of the following: never married, Þrst marriage, divorced once (not re-married),
second marriage, second divorce, third marriage, and third divorce. In each period,
individuals are presented with options in each career sector and potential marriage
offers in the marriage market. The probability of receiving a marriage offer in any
period is conditional on current and previous marriage and career decisions and personal
characteristics (AFQT score, unobserved type, etc.). Wage offers in each occupation
are also conditional on current and previous marriage and career decisions and personal
characteristics. In each period, individuals maximize their expected lifetime utility by
deciding whether or not to change their current marital status (conditional on receiving
an offer), and whether to work in one of the occupations, go to school, or stay at
home.
Estimation of the model involves the numerical solution of a Þnite-horizon, discrete-
choice dynamic programming problem, nested within an algorithm that maximizes a
likelihood function. The dynamic program is solved by backward recursion, and the
likelihood function is constructed by simulating the solution of the optimization prob-
lem for a set of artiÞcial agents, and then maximizing the probability that the simulated
agents match the decisions of individuals in the data. This method is particularly suit-
able for data sets with missing endogenous state variables as is the case when using the
NLSY, especially since the NLSY was conducted only every other year after 1992.
3
Although there is considerable research on female marriage and labor supply de-
cisions (Heckman and MaCurdy (1980), Hotz and Miller (1988), Eckstein and Wolpin
(1989), van der Klaauw (1996), and many others), the interaction of marriage and career
decisions for men has received scant attention. A few recent papers have developed
structural models of two-sided matching in the marriage market (Brien, Lillard, and
Stern (2002), Seitz (2002), and Wong (2002)), but have primarily focussed on the role
of the endogenous sex ratio in determining matches and how cohabitation interacts
with divorce and match quality. These papers do not address the dynamic marriage
and career decisions of men, or estimate the returns to career decisions in the marriage
market and the returns to marriage decisions in the labor market.
A few of the issues addressed in this paper have been examined in isolation within a
non-dynamic framework. In particular, the marriage premium has received considerable
attention, as the literature has tried to sort out whether married men earn more because
of any causal relation or whether the premium represents a high degree of positive
sorting into marriage (see Korenman and Neumark (1991), Cohen and Haberfeld (1991),
Gray (1997), and Hersch and Stratton (2000)). However, the literature has never
modeled the decision to get married in order to control for endogeneity and individual
heterogeneity. There is also signiÞcant work on the issues of divorce and re-marriage.
Becker, Landes, and Michael (1977) examined marital turnover and suggested two
general causes: (1) search is costly and meetings occur randomly, therefore, a marriage
offer which was accepted in the past may be discarded when a better match is offered in
the future; and (2) traits that determine the gains from the marriage match can change
in unpredictable ways, such as changes in labor market prospects. Becker, Landes, and
Michael (1977) use cross-sectional data to examine these issues, and Weiss and Willis
(1997) build on their analysis by using longitudinal data to try and separate changes in
match quality from initially bad matches. However, Weiss and Willis (1997) focus
on the determinants of the dissolution of the Þrst marriage only. This paper builds
on this literature by explicitly modeling the dynamic search process in the marriage
market, while embedding a model of marital turnover in a model of career decisions.
The model of career decisions closely parallels the framework developed by Keane and
4
Wolpin (1997). This paper enriches their model of occupational choice by seeing how
occupational choice interacts with marriage decisions.
2 The Data
The data are taken from the National Longitudinal Survey of the Labor Market Expe-
rience (NLSY 1979) from 1979 until 1998. The sample consists of a random sample
of white men ages 14-21 in 1979 taken from the core NLSY sample. We follow each
individual from the time that they are at least 16 years old until the end of the sample
period (up to 39 years old), and record the individuals marriage, education, work, and
occupational status for every calendar year available. The main analysis consists of a
sample of 2,155 men who report their joint career and marital status for at least three
years. Individuals appear in the sample for an average of 15.4 years.1 In addition, the
analysis uses data on weekly wages for individuals who work, as well as measures for
each respondents mental ability (AFQT score adjusted for age at which the test was
taken) and parental background (whether both parents were living with the respondent
at age 14 and the education level of each parent).2
The estimation method used in the analysis does not require each individual to
have complete retrospective data for any given age to be included in our sample, as is
done in many previous dynamic models such as Keane and Wolpin (1997). Therefore,
the estimation strategy will be able to use information on each individual at each point
in time, regardless if the persons wage, marital or career status is unknown at any given
time period.
I now describe how individuals were categorized into one marriage category and
one occupation sector for every year where sufficient information was non-missing.
1Ninety percent of the men in the sample report for at least for 10 years, and sixty-three percentfor at least 15 years.
2The wages for self-employed workers were treated as missing. The sample was restricted to thosewith non-missing AFQT scores and family background characteristics (whether both parents wereliving with the respondent at age 14 and the education level of each parent).
5
2.1 Marriage Sectors
Each individual in the NLSY was asked about their marital status each sample year
and were asked retrospective questions about when each of their marriages started and
ended. When sufficient marriage information was available, individuals were classi-
Þed into one of seven marriage sectors in each year: (1) single (never married), (2)
Þrst marriage (currently married for the Þrst time), (3) Þrst divorce (married once,
divorced, and currently single), (4) second marriage (divorced once and remarried),
(5) second divorce (divorced twice and currently single), and (6) third marriage
(re-married for the second time), and (7) third divorce (divorced for the third time
and currently single). Table 1 displays the marriage sector distribution for the sample
over time. As expected, individuals gradually get married, divorced, re-married, etc.
The median individual waits until the age of 25 to get married, but there are signiÞcant
differences according to the persons AFQT score. For those in the bottom 20 percent
of the AFQT distribution, the median age at Þrst marriage is 24 while the those at the
top 20 percent marry at the median age of 27. Overall, 71 percent of Þrst marriages
last at least 10 years, but again, there are sharp differences in the marriage survival
rate according to AFQT scores: 84 percent of Þrst marriages survive at least 10 years
for men at the top 20 percent of the AFQT distribution, while only 59 percent are still
active for those at the bottom 20 percent. Clearly, AFQT score is positively associated
with age-at-Þrst marriage and more durable marriages.
2.2 Career Sectors
Individuals were classiÞed in each period into one of the following four mutually exclusive
career sectors: schooling, blue-collar, white-collar, and home. These classiÞcations
are similar to those used by Keane and Wolpin (1997). To be considered working in
either the white-collar or blue-collar sector, the individual had to work at least 30 weeks
out of the calendar year for at least 20 hours per working week. To be classiÞed as
being in school, the individual had to complete a year of schooling during the calendar
6
year.3 Respondents with non-missing information about weeks worked and schooling
status who were not classiÞed into one of the other three sectors were placed into the
home sector.
Table 2 reveals the expected pattern of career choices as the cohort ages over time.
At Þrst, most of the sample attends school and gradually they move into the blue-collar
and white-collar occupations over time. Interestingly, the data displays a cross-over
pattern as workers who start out in blue-collar tend to move over to the white-collar
sector over time a pattern not picked up by Keane and Wolpins (1997) study of career
choices since their study stopped at the age of 26. There are also very stark differences
in career patterns across levels of AFQT: men with lower AFQT scores typically work
less (the home sector), study less, and tend to work in the blue-collar sector more
than the white-collar sector.
2.3 The Interaction of Marriage and Career
A Þrst pass through the data reveals a strong and complex relationship between marriage
and career outcomes. The Þrst column of Table 3 presents the results of a standard
wage regression for the core sample of male workers in 1992. After controlling for a
wealth of personal characteristics (AFQT, mothers education, etc.), wages are shown
to be very signiÞcantly related to the occupational choice of workers and their marital
status. White-collar workers earn a 17% premium while married workers earn 19%
more than blue-collar workers. Interestingly, divorced workers earn less than married
workers, and are not signiÞcantly different than single workers. These patterns persist
if wage regressions are run separately for white-collar and blue-collar wages, as shown
3SpeciÞcally, the individual had to satisfy three requirments to be in the schooling sector in anygiven calendar year. First, the respondent must report that Highest Grade Completed as of May 1has increased by one since the previous calendar year (using the *created Highest Grade Completedby May 1 of Survey Year variable). Second, the respondent must satisfy at least one of the followingtwo items: (a) report to be enrolled in school as of May 1 of the calendar year (using the *createdEnrollment Status as of May 1 of Survey Year variable) or (b) report that the main activity duringthe interview week in the calendar year as going to school (using the *created Employment Statusvariable). Third, the respondent must satisfy at least one of the following two: (a) worked less than24 weeks during the calendar year, or (b) worked more than 24 weeks but less than an average of20 hours per working week. In addition, for the year 1979, individuals who were currently in thegrade appropriate for someone who went continually to school were classiÞed retrospectively as beingin school from the age of 16 to the current age.
7
in the last two columns of Table 3. Surprisingly, married workers earn more in blue-
collar (22%) than white-collar (14%) compared to non-married men. Again, divorced
men are not signiÞcantly different from single men. These Þndings, however, are
just correlations and do not establish any causal relationships. In particular, the large
marriage premium, which is a common result, may not be due to the effect of marriage
on wages, but rather may result from a non-random selection of men who choose to get
married.
Table 4 shows how marital status is strongly correlated with career choices. Con-
ditional on wages, blue-collar workers are signiÞcantly less likely to be married than
white-collar workers and more likely to get divorced if they do get married. Again, the
direction of causality is not clear. Blue-collar workers may have more trouble Þnding
a good match in the marriage market, thus lowering their probability of being married.
When they do Þnd a match, the quality of their match may be lower, thus increasing the
likelihood of getting divorced. On the other hand, a non-random selection of less able
and less stable men into the blue-collar sector could also account for these correlations.
A similar story can be told about individuals in the home sector, who are also less
likely to be married and more likely to get divorced. In contrast, white-collar workers,
conditional on wages, take longer to get married but are more likely to stay married.
This pattern may result from the idea that these workers can afford to wait until they
Þnd a high quality match, which in turn, produces a more durable match in the long
run (see Bergstrom and Bagnoli (1993), Weiss (1997), and Weiss and Willis (1997) for
similar explanations). Higher wages, on the other hand, increase the chance of being
married and staying married. But, again, this correlation may result from unobserved
characteristics which may make men earn more and attract higher quality matches in
the marriage market.
In general, it is not clear whether the characteristics of each sector cause people
to behave in certain ways, or whether the self-selection of certain types of people into
certain sectors is responsible for the strong correlations between marriage and career
choices in the data. The model described in the next section will enable us to untangle
these complicated relationships by estimating the underlying structural parameters of
8
the decision making process which generates the observed marital and career outcomes
in the data. Although many of the issues raised above have been addressed in the
existing literature, they have generally been analyzed in isolation. There is no study that
formulates and estimates a model that can explain labor market and marital outcomes
for men within a single uniÞed framework.
3 The Model
This section presents the basic structure of the model and the parameterizations of
each structural equation. The solution to the model and the estimation method is
also discussed. The model corresponds to the decision problem of a single individual
choosing his career and marital status in each time period t (t = 1, ..., T ) in order
to maximize his expected present discounted value of available alternatives which are
based upon previous decisions. Each period is associated with a certain age (ages
16-39). In addition, each individual enters the Þrst period with initial background
variables consisting of: being single, afqt (corresponding to quintile of the individuals
age-adjusted AFQT score, so that afqt ∈ 1, 2, 3, 4, 5), and family (family backgroundcharacteristics based on whether both parents lived with the respondent at age 14 and
the schooling levels of each parent).4 There is also unobserved heterogeneity in men,
characterized by three different types of men (type ∈ 1, 2, 3).
3.1 Marriage and Career Choice Set
In each period, individuals choose one of four broadly deÞned career sectors: home
(nonemployment) (kt = 0), school (kt = 1), blue-collar employment (kt = 2) and
white-collar employment (kt = 3). The number of years accumulated in each career
4This paper follows Taber (2001) by assuming that AFQT (adjusted by the age at which the testwas taken) is an exogenous variable. This does not mean that AFQT is equivalent to IQ or that itis inherited genetically. AFQT is simply regarded as a measure of ability at the time the test wastaken. In addition, I do not Þnd that the AFQT score is endogenous to education level at the timeof the test which suggests that AFQT is comparable across all ages in the sample. I tested this byregressing the age-adjusted AFQT score on the eventual highest grade completed and found that therelationship between AFQT and eventual schooling was stable across all age levels. Also, in order touse AFQT (which is a continuous variable) in the state space, the score was discretized into quintilesand the score associated with each quintile is the value of a standard normally distributed variable inthe middle of the quintile.
9
choice k at the end of year t is represented by xkt (i.e.Ptτ=1
P3k=0 xkτ = t). Initial
conditions for the experience levels in each sector are normalized to zero: x10 = x20 =
x30 = 0 . Individuals are free to choose any career sector in any given period.
In order to capture the logical sequence of marriage possibilities, the marital choice
set contains ten options mt:
Single, Never Married mt = 0First Marriage, Type 1 wife mt = 1First Marriage, Type 2 wife mt = 2Divorced Once, Single mt = 3Second Marriage, Type 1 wife mt = 4Second Marriage, Type 2 wife mt = 5Divorced Twice, Single mt = 6Third Marriage, Type 1 wife mt = 7Third Marriage, Type 2 wife mt = 8Divorced Three Times, Single mt = 9
The initial condition for marital status is never being married (m0 = 0). However,
unlike career choices, marriage decisions are restricted to choosing among the available
options in each period. The set of potential options includes keeping the same marital
status from one period to the next:
mt = mt−1 if mt−1 ∈ (0, 1, 2, 3, 4, 5, 6, 7, 8, 9)
Or, individuals can potentially change their marital status according to the fol-
lowing logical sequence of marriages:
mt ∈ (1, 2) if mt−1 = 0mt ∈ (3, 4, 5) if mt−1 ∈ (1, 2)mt ∈ (4, 5) if mt−1 = 3mt ∈ (6, 7, 8) if mt−1 ∈ (4, 5)mt ∈ (7, 8) if mt−1 = 6mt = 9 if mt−1 ∈ (7, 8)
In this manner, marital options are restricted so that marriages must occur in
sequential order (i.e. they cannot go directly from their Þrst marriage to their third
marriage). In addition, individuals are potentially free to go from one marriage to the
next marriage without spending a period being divorced and single, and individuals
are allowed to marry one type of wife in one marriage and a different type in a future
marriage (i.e. mt−1 = 4 and mt = 8).
10
Within the set of potential options described above, the only option that is always
available is the option to be single. That is, single men can always remain single
(mt = mt−1 ∈ (0, 3, 6, 9)), and married men always have the option to divorce theircurrent wife and become single again.
However, available marital options are restricted in two further ways: (1) marriages
can only occur if a man receives an offer, and (2) marriages may be terminated by the
wife (i.e. the man does not have the option to continue with his current marriage).
Formally, a married man will have his current marriage terminated exogenously by his
wife with probability πDt , in which case the man has to be single in the next period or
get re-married if a new marriage offer was forthcoming. Marriage offers arrive with
probability πµt , which is the probability of receiving an offer to marry a woman of type
µ (µ = 1 or 2). Marriage offers can be received by single or married men, and men are
always free to reject the offer in the hope of getting a better offer in the future. If no
marriage offer is forthcoming, a single man must remain single and a married man can
remain married (if his wife did not terminate the marriage) or become divorced and
single again.
Therefore, the set of available marital options at each time t, denoted by Mt, are
a function of the marital status in the previous period, mt−1, and are conditional on
whether a new marriage offer is received, the type of marriage offer, and whether the
wife terminates the current marriage (if he was married).
The full choice set thus contains four career and ten potential marriage options,
all of which are assumed to be mutually exclusive and exhaustive. However, not all of
the options in the choice set have observable counterparts in the data. The four career
choices are observable as well as the marital status of each man in our sample, including
whether he is divorced or on his Þrst, second, or third marriage. In contrast, the wifes
type is not observable in the data, but one can think of heterogeneous wife types as
representing variation in the quality of marriage matches, which is partially observable
in the data because we do observe that certain types of marriages are more successful
(endure longer) and are able to withstand certain types of observable shocks (i.e. wage
11
shocks).5 Also, heterogeneity in wife types allows for a non-trivial search process in
the marriage market. Without variation in the quality of wives, the search process for
wives in the marriage market would simply degenerate into each man marrying the Þrst
woman he meets. With heterogeneity in wives (or match offers), the model allows for
the possibility that a man would turn down an offer to marry a certain type of woman in
the hope of receiving an offer to marry a different type of woman in the future, and thus
captures the complex, forward-looking decision-making process inherent in the choice
of marital status. As described later, this strategic decision process in the marriage
market will interact with strategic career choices.
3.2 Parameterizations
3.2.1 Marriage Utility
The single-period utility associated with marital status mt is a function of whether the
person is married, the type of wife (if he is married), and how many divorces the person
went through. The latter element is intended to capture the alimony (monetary costs)
and psychological costs that accompany a divorce. After each divorce, the individual
is assumed to pay an additional monetary cost each period thereafter. Thus, divorce
costs in period t can be characterized by the following:
div(mt) = 0 if mt ≤ 2= δd1 if 3 ≤ mt ≤ 5= δd1 + δ
d2 if 6 ≤ mt ≤ 8
= δd1 + δd2 + δ
d3 if mt = 9
(1)
The current utility of being married also depends on whether the person is married
to a type 1 wife (mt ∈ 1, 4, 7) or type 2 wife (mt ∈ 2, 5, 7). Thus, the model will allowfor the pecuniary returns (the wifes income) and the monetary equivalent of the utility
of being married to each type of wife to differ. Formally, this is represented by:
marr(mt) = δm1 + εmt if mt ∈ 1, 4, 7 (2)
= δm2 + εmt if mt ∈ 2, 5, 8
5It does not necessarily have to be the case that all men agree who are the good types of potentialwives (matches) and who are the less-good types of potential wives. All that is necessary is that peopledo believe that their marriage quality will be different with different women.
12
= 0 otherwise (single)
where εmt is an identically distributed random utility shock to the current marriage,
which is uncorrelated over time, and is only experienced by individuals who choose to
be married. This speciÞcation allows for variation in the utility of marriage across types
of wives (variation in match quality), and also for the utility of the marriage itself to be
stochastic, since as pointed out in Mortensen (1988), the quality of the marriage match
is revealed and can change over time.6 Therefore, if the marriage match deteriorates
by the individual receiving a sufficiently bad shock to the marriage, he may decide to
terminate the marriage and become single or marry someone else if a new marriage offer
was received.
The net current-period utility associated with marital status mt is the combined
marriage utility and divorce costs:
um(mt) = marr(mt, εmt ) + div(mt) (3)
3.2.2 Career Utility
The current period utility associated with each of the four career sectors is dependent
on the marital status of the individual (mt), the accumulated levels of experience in
each career sector as of year t (x0t, x1t, x2t, and x3t), .the individuals type (type = 1, 2,
or 3), and the individuals afqt quintile. To ease the notation, the vector of experience
levels in all four career sectors is represented by Xt. In general, the one period utility
of choosing career sector j is represented by ukt (kt,mt),which are speciÞed for each of
the four career sectors.
Home Sector Utility (kt = 0):
ukt (0,mt) = b0 (type, afqt, t) + εk0
t (4)
where b0 is the one-period non-monetary value of leisure, which depends linearly on a
quartic of the individuals age t, type, and afqt score. The εk0t term is a stochastic
6Another source of wife heterogeneity could stem from heterogeneity in bargaining power (Chi-appori, Fortin, and Lacroix (2002)).
13
shock to the value of leisure in period t which is uncorrelated over time. The structure
of all the shocks in the model will be discussed later.
Schooling Sector Utility (kt = 1):
ukt (1,mt) = b1 (t) + tuition(x1t) + entry
1(x1t, kt−1) + εk1t (5)
where b1 is the one-period net utility of being in school, taking into consideration both
the direct monetary costs of schooling and the potential consumption value of schooling.
The total net utility b1 varies according to the individuals age, represented by t. The
tuition function allows for the costs of schooling to change with levels of schooling and is
parameterized as a step function with steps for high school (x1t ≤ 2), college (x1t ≤ 6),and graduate school (x1t > 6). The entry1 function allows for the one-time costs of
returning to school from a different sector (i.e. kt−1 6= 1) to vary with the amount
of schooling. That is, the costs of returning to high school are γ11,while returning to
college or graduate school costs γ12.
Blue-Collar Sector Utility (kt = 2) :
ukt (2,mt) = wage2(Xt, mt, type, afqt, kt−1, ε
k2t ) + entry
2(Xt, kt−1) (6)
where wage2 is the blue-collar wage offer, and entry2 is the one-time (non-wage) cost
of entering the blue-collar sector if the individual was not working in the blue-collar
sector in the previous period (kt−1 6= 2). This non-wage entry cost captures the ideathat there may be search costs in Þnding the blue-collar offer or starting to work in
the blue-collar sector (transportation, clothing, etc.). The wage offer in the blue-collar
sector is a Mincer-like wage function which depends on the accumulated experience
in each career choice Xt (completed schooling, blue-collar experience, and white-collar
experience), the individuals afqt score, the individuals type, and the individuals age
t. Inclusion of a return to white-collar experience in the blue-collar sector allows for the
full or partial transferability of occupation-speciÞc experience in the other occupation.
Some of these variables are interacted with each other in a typical log wage speciÞcation.
14
In addition, the wage offer is a function of marital status, where the state of being
married affects the intercept of the wage offer, and it does so differentially depending on
the type of current wife. The state of being married is also allowed to affect the returns
to accumulated experience in the sector, thus allowing for marital status to affect not
only the intercept but also the experience proÞle of the wage function. Furthermore, the
speciÞcation allows for changes in the intercept and experience proÞle for individuals
who decide to divorce, so any returns to being married could dissipate upon the
transition into divorce.
There is also a return to continuing to work in the blue-collar sector (kt−1 = 2)
which, along with the entry cost function, is designed to capture the persistence of career
choices across time periods (see Keane and Wolpin 1997). Also, including a return to
staying in the same sector is equivalent to incorporating a human capital depreciation
effect. The log wage offer is also subject to a linear stochastic component εk2t , which is
uncorrelated across time.
White-Collar Sector Utility (kt = 3) :
ukt (3,mt) = wage3(Xt, mt, type, afqt, kt−1, ε
k3t ) + entry
3(Xt, kt−1) (7)
where each component is deÞned analogously to the utility components in the blue-
sector, although the parameters differ for each sector.
3.2.3 Correlation of Marriage and Career Shocks
In each period t, an individual receives four separate shocks (as shown above) to each
career sector ( εkjt , j = 0, 1, 2, 3), and if they are married, a marriage shock εmt . All
of these shocks are presumed to normally distributed (with mean zero) and contempo-
raneously correlated, but mutually serially independent over time (Keane and Wolpin
(1994)).
3.2.4 Marriage Offer Functions
In each period t, individuals may receive a new marriage offer with a type 1 or type
2 wife. The probability of receiving an offer of either type is speciÞed as a tri-variate
15
logit where the three outcomes are: (1) no offer, (2) an offer from a type 1 female, or
(3) an offer from a type 2 female. The probability of receiving an offer from a type µ
woman is:
πµt = πµ(Xt−1, kt−1, mt−1, type, afqt, t), µ = 1, 2
where the base (suppressed) state is not receiving an offer at all. The probability
of receiving an offer from either type of female depends on the individuals historical
and current career and marital choices, as well as the individuals age, type, and afqt
score. Note that married people are allowed to receive new offers, but the probability
of receiving an offer is likely to be affected by marital status. Individuals can only get
married if they receive an offer, and their choices are restricted to the type of woman
who is giving the offer.
3.2.5 Exogenous Divorce (Involuntary Termination of the Marriage by theWife)
Individuals who are currently married may have their marriages terminated unilaterally
by their wives. The probability of an exogenous divorce at time t is speciÞed as:
πdt = πd(kt−1, x1t−1, afqt, type)
Thus, the exogenous termination of marriages is modelled as a function of the
last periods career choice, accumulated levels of schooling x1t−1, afqt score, and type.
These variables capture the extent to which divorce probabilities depend on permanent
elements of the persons potential wage earnings (occupation, education level, AFQT,
and type). Although the persons afqt and type are assumed to be exogenously deter-
mined for each individual, each person can affect his divorce probabilities through the
choice of occupation and schooling investments over time.
3.2.6 Type Probabilities for Men
Each man is assumed to be one of three discrete types corresponding to three mass points
in a non-parametric distribution of permanent unobserved heterogeneity (Heckman and
16
Singer (1984)). The probability of being a certain type (type = 1, 2, 3) of male is
modeled as a tri-variate logit:
πtype = πtype(afqt, family), type = 1, 2, 3 (8)
where the probability of being certain type depends on the individuals afqt score and
family background, all of which are assumed to be exogenously determined for each
individual.
3.2.7 Objective Function
The individual is assumed to maximize the present discounted value of lifetime utility
from age 16 (t = 1) to age 39 (t = T ). Let Ωt represent the relevant information
set with which the individual enters period t. Ωt includes the individuals history
of career decisions (denoted by Xt−1), marriage decisions (inferred from mt−1) and
the individuals type, afqt score, and family background. Given this set of relevant
information, the one-period utility associated with any combination of marriage status
mt and career choice kt is denoted by U(mt, kt | Ωt), and is determined by equations(3) and (4)-(7) above:
U(mt, kt | Ωt) = ukt (kt,mt) + umt (mt) (9)
This speciÞcation clearly shows the interaction between marriage and career: cur-
rent and historical marital decisions affect the utility of career choices by affecting the
wages in each occupation, while current and historical career choices affect marriage
opportunities by affecting the chances of getting a marriage offer, the type of offer, and
the probability of receiving an exogenous divorce in the current and future periods.
Thus, the interaction of marriage and career choices demands a maximization decision
based on the joint and forward-looking marriage and career choice path.
Although all four career choices are available each period, marriage options are
restricted to the available set of marital options deÞned by Mt above. Therefore, the
available choice set in period t is given by the Cartesian product of the four career
sectors multiplied by the marriage options contained in Mt. We denote the choice of
17
element j in this feasible set of joint marriage and career choices in period t as djt = 1
(j = 1, ..., Jt), and the utility associated with that choice as Ujt (speciÞed in equation
(9)). The individuals objective function is then represented as:
Vt(Ωτ ) = maxdjt
E
TXτ=t
JτXj=1
δT−τU jτ djτ | Ωτ
(10)
where δ is the discount factor (Þxed at 0.95) and E is the expectation operator taken
over the joint distribution of utility and marriage shocks (εk0t , ε
k1t , ε
k2t , ε
k3t , ε
mt ), as well as
the distribution of marriage offer probabilities (π1t and π
2t ) and exogenous divorce prob-
abilities πdt . The solution to this problem yields the optimal stream of joint marriage
and career decision over time.
3.3 Model Solution and Estimation
3.3.1 Solution
The solution of the model is not analytic, and therefore, is solved numerically using
backward recursion starting from a terminal age T . The maximization problem in
equation (10) can be re-written as the maximization over the value functions of the
available set of joint marriage and career states j (j = 1, ..., Jt) at time t, denoted as
V jt (Ωτ ), which satisfy the Bellman (1957) equation:
Vt(Ωt) = maxhV 1t (Ωt), . . . ,V
Jt (Ωt)
i
V jt (Ωt) = Ujt. + δE
hVt+1(Ωt+1) | (djt = 1), Ωt
i(11)
Therefore, given any set of parameters, solving the model consists of simulating
all of the stochastic components of the model at each point in the state space (every
possible combination of historical marriage and career decisions for every type and afqt
score up to period t), and using backwards recursion to calculate U jt and E [Vt+1(Ωt+1)]
(see Keane and Wolpin (1994, 1997)). The latter term is called the Emaxt+1 function
for convenience. At each iteration in the estimation, 30 draws of the entire set of
stochastic components were taken according to the current set of parameters to estimate
18
the Emaxt at every point in the state space. The value of EmaxT for the terminal
period T is parameterized as function of the individuals afqt and historical and terminal
state choices: EmaxT (XT ,mT , kT , afqt).
3.3.2 Estimation
To estimate the model, the numerical solution of the dynamic programming problem
described in the previous section is nested within an algorithm that maximizes a like-
lihood function. The likelihood function is constructed by simulating a set of choice
histories and matching these choice histories with observed choices in the data. The
estimation strategy deals effectively deals with the problem of unobserved initial con-
ditions (see Heckman (1981)) and state variables. These problems can be quite severe
when constructing marriage and employment histories from NLSY data, since the rele-
vant information is frequently missing for some respondents in various years, especially
after 1992 when the NLSY survey was conducted every other year. Therefore, this
technique allows us to use all the marriage and career choice information contained in
the data whenever the data is available, without having to worry about constructing the
complete marriage and career history for each respondent.7 Consequently, the sample
size in this analysis is much larger than the samples used in many previous dynamic
programming studies using the NLSY (such as Keane and Wolpin (1997)).
The estimation algorithm is based on simulating the complete marriage and career
histories of a set of artiÞcial agents (n = 1, . . ., N). Given a set of parameter values,
the simulation for agent n is performed as follows:
1. Draw agent ns parental background (family) and afqt according to the actual
proportions in the data (and according to the actual correlations of these variables
to each other in the data - which were estimated by logits and multivariate logits
outside the estimation algorithm using the NLSY sample).
2. Using the simulated background variables (family) and afqt, draw the agents
type (type = 1, 2, 3).7The estimation uses techniques developed by a long list of papers. A partial list includes Heckman
(1981), Heckman and Singer (1984), Rust (1987), Hotz and Miller (1988), Keane and Wolpin (1994,1997, 2001), and Hotz, Miller, Sanders, and Smith (1994).
19
3. Draw from all the stochastic elements in the model (εk0t , ε
k1t , ε
k2t , ε
k3t , ε
mt ), as well
as from the marriage offer functions (π1t and π
2t ) and exogenous divorce proba-
bility function (πdt ) to determine the available marriage options and career offers
(conditional on ns afqt and type).
4. According to the agents type, afqt, and realizations of the stochastic elements in
step (3), the agent considers the Emaxt+1 term for each option, which was already
constructed at the current parameterization, and evaluates each current marriage
and career option using equation (11). The agent chooses the joint marriage and
career option with the highest expected value over the current set of J options:
maxhV 1t (Ωt), . . . ,V
Jt (Ωt)
i5. The state variables (Xt, mt, kt) are updated according to the choice in step (4).
6. Repeat steps (3)-(5) until t = T .
Doing this N times produces N artiÞcial agents with a complete set of marriage
and career outcomes over T periods (N = 80, 000 in the actual estimation). The
likelihood function is then built using a frequency simulator, although as Lerman and
Manski (1981) point out, the probability that the entire career and marriage choices
of a simulated agent (including wages) matches someone in the data is very inÞnitely
small. Instead, the likelihood is built on the simulated frequencies of period-by-period
choices.
In particular, there are 28 possible joint career and marriage choices (four career
choices and seven marriage states) observed for every individual in the data at time
t. Denote the probability that individual i, given his exogenously determined type and
afqt score, chooses the joint marriage and career sector j (j = 1, ..., 28) and earns log
wage wijt at time t :
P ittype,afqt(choiceit = j, lnwage=wijt | type, afqt) =
prob(choiceit = j | type, afqt) ∗ ϕÃÃwijt − µjtσjt
!| choiceit = j, type, afqt
!(12)
20
where ϕ(·) is a standard normal pdf and µjt and σjt are the mean and standard deviationof log wages conditional on choosing the joint marriage-career state j at time t and
given the individuals type and afqt. The Þrst component can be estimated by the
proportion of simulated agents with the same type and afqt who choose j at time t.
The mean and variance of the conditional wage distribution can be estimated similarly
by computing the mean and variance of log wages for simulated agents with the same
type and afqt who choose j in time t. Substituting these estimates into equation (12)
yields the simulated probability bP ittype,afqt. The probability of the sequence of person ischoices and wages over time, given type and afqt,is then estimated as:
prob(choices i | type, afqt) =TYt=1
bP ittype,afqt (13)
If the persons type were known, equation (13) would be the likelihood contri-
bution of person i. However, a persons type is unknown, but can be estimated as a
function of observable initial characteristics (family and afqt). Therefore, the likeli-
hood contribution for person i is simply the weighted sum of equation (13) over the
three possible types using the probability of person i being each type as the weight:
Li =3X
type=1
prob(choices i | type, afqt) ∗ πtype(i = type | family, afqt) (14)
where πtype is deÞned above in equations (8) as the probability of person i being a
certain type (type = 1, 2, 3) conditional on afqt and family. Thus, we have derived
the contribution of each person in the data to the likelihood function conditional on his
afqt score and family background characteristics.8 In practice, the simulated likelihood
function was optimized using a simplex algorithm since the likelihood function is not
smooth, and standard errors were computed by estimating the derivatives of the log
8Although it is not obvious, the above estimation strategy is a special case of Keane and Wolpin(2001) when their classiÞcation error rates are set to zero. This is true only if there were no wages toestimate.
21
likelihood function with respect to each parameter at the maximum.9
The likelihood in equation (14) is composed only of elements when there is a
recorded choice for person i in the data. In periods where person i does not report a
choice or wage, this element is simply dropped in the process of building the likelihood.
In this manner, we simply match whatever choices are reported in the data, conditional
on exogenous initial conditions (family and afqt), to simulated moments (which are
also conditioned on the same initial conditions as state in step (1) above) based on the
current set of parameters in the iteration process. In theory, we could match person
is choices in each period t to simulated moments conditional on person is previous
choices. However, this would force us to throw out any observation in the sample
where the previous periods choice is unobserved. As a consequence, the sample size
would be dramatically reduced, in particular, because the NLSY was sampled only every
other year from 1992 onward. Thus, even if we conditioned on the previous periods
choice, the sample used in the analysis would contain few people beyond the age of 29
and no one beyond the age of 33. Since many important career and marriage decisions
are taken by men in their 30s, I decided to build the likelihood in such as way as to be
able to explain and Þt the marriage and career paths of men well into their 30s.
4 Results
4.1 Fit of the Model
Before discussing the results, I Þrst present the Þt of the model. Figures 1a and 1b
show that the model produces patterns of career choices very similar to those reported
in the NLSY sample. In particular, the model captures the sharp drop in schooling
and the eventual overtaking of the blue-collar sector by the white-collar sector. Figure
2 displays a close Þt of the model to blue-collar and white-collar wages as well, in
particular during the prime ages of working. The model Þt of the marriage categories
9Standard errors were estimated as follows. Let gi be the vector of derivatives of the log contributionto the log likelihood of person i with respect to the set of parameters θ : gi = ∂ lnLi
∂θ . This derivativewas approximated by taking small steps in the estimated parameters θ : bgi = lnLi(θ)−lnLi(θ+h)
h where
h = θ ∗ 10−3. The covariance matrix is then estimated by:³P
i bgibg0i
´−1
.
22
is shown in Figures 3a to 3d. The model picks up the general patterns in marriage
behavior quite well. Figure 3a shows a very close Þt to the age-at-Þrst marriage and
the stock of men who are currently on their Þrst marriage. Concerning second and
third marriages, the model picks up the broad patterns and magnitudes, but the Þt is
clearly not exact (the very small scale on these graphs exaggerate the appearance of
rather small differences).. In Figure 3d, the model is shown to estimate the survival
rates of Þrst marriages quite well, although marital success is somewhat overestimated
beyond six years of marriage.
The model Þt according to AFQT levels are depicted in Figures 4a to 4d. The
model captures very closely all of the differences in the marriage and career patterns
across levels of AFQT scores. In particular, the model shows how AFQT scores are
positively related to schooling, working in white-collar, higher wages, delays in marriage,
and higher survival rates of the Þrst marriage. Most notably, the model estimates
that 42% of marriages fail by the tenth year for men in the lower quintile of the AFQT
distribution, while the actual number in the NLSY is 41% (Table 4b). In stark contrast,
the model predicts that a much lower 12% of marriages will fail in the Þrst ten years
for men in the upper AFQT quintile while the actual number is 16% in the NLSY.
Therefore, the model captures not only the broad marriage and career patterns for all
men over time, but also matches the cross-sectional variation in these patterns across
men with different AFQT scores.
The fact that the model picks up the survival rates of marriage well indicates that
the model is able to Þt moments in the data that the likelihood is not explicitly trying
to Þt. The likelihood is built to Þt the proportions of men who choose each state in each
period, but information concerning transitions across sectors is not entered directly into
the likelihood function. Therefore, a good test of the model is to see how the model Þts
moments which the likelihood function is not explicitly trying to match. The model
appears to do this quite well by Þtting the survival rates of marriage (Tables 3d and 4b),
and by Þtting the transitions into and out of the various marriage and career sectors.
These transitions are displayed in Tables 5 and 6, which show a very close match to the
transitions into and out of almost every marriage and career sector. Not surprisingly,
23
the transitions into and out of the states with the most observations (never married,
Þrst marriage, school sector, blue-collar, white-collar, etc.) are matched very well, while
the transitions from states with much less data (second divorce, third marriage, home
sector) are matched less well.
Overall, the model picks up the marriage and career patterns for the overall sample
over time, and the cross-sectional variation according to levels of AFQT. The model
also produces a close match to the transitions into and out of marriage and career states,
which indicates that the estimation procedure is able to Þt moments that the likelihood
is not explicitly trying to Þt which should increase our conÞdence in the models
predictions. However, it is left to the reader to decide whether the Þt is good enough
to draw credible inferences from the results and counterfactual experiments presented
in the rest of the paper.
4.2 Discussion of the Estimates
There are too many (155 to be exact) coefficients to discuss individually, but sev-
eral overall patterns emerge in the estimates presented in the Appendix. First, there
are three types of men which are distinctly characterized by increasing quality of the
parental background (parental education and whether the person lived with both par-
ents at age 14). Type 1 men have the most troubled family background, while Types 2
and 3 have increasingly better backgrounds. Not unsurprisingly, Type 3 men typically
earn higher wages in both the white and blue collar sectors, while Types 2 and 1 lag
behind in that order (although Type 2 men actually have a slightly higher blue-collar
intercept than Type 3 men). Types 1 and 3 work predominantly in the white-collar
sector while Type 2 men tend to pursue their comparative advantage in the blue-collar
sector.
The estimates also show a broadly consistent pattern whereby higher investments
in human capital (in education or work experience) increase the probability of getting a
marriage offer and decreasing the chance of suffering an exogenous divorce. In addition,
these same variables increase the chances of receiving a better marriage offer with a
Type 2 wife the utility of Type 2 wives is roughly double that of Type 1 wives. In
24
particular, men who go to college signiÞcantly increase their chances of marrying a
Type 2 woman. Figures 5a and 5b illustrate the patterns of marriage types in the
aggregate and according to AFQT levels. Most notably, marriages to Type 2 wives
(the better type) occur only after the age of 21 while marriages to Type 1 wives occur
earlier, thus suggesting that Type 2 wives are likely to be college educated or high ability
women. Also, Type 2 wives seem to be in the minority, with twenty percent of the
men married to Type 2 wives and Þfty percent to Type 1 wives although higher levels
of AFQT are associated with more marriages to the better Type 2 wives (a form of
high assortative mating as found by Fernandez, Guner, and Knowles (2001) in many
advanced countries).
Estimates for the exogenous divorce probabilities also show that higher levels of
ability and education reduce the probability of an exogenous marriage termination.
Interestingly, divorce probabilities are noticeably higher in the home sector than the
other three career sectors.
Overall, the estimates show that higher levels of ability and education increase
the chances of getting married, Þnding a higher quality match, and staying married
longer. These patterns demonstrate the interactive, joint, and forward-looking nature
of the marriage and career decision-making process for men.
Marriage Premium
The estimates of the marriage premiums contrast sharply with those using OLS
in Table 3, which were 22% for blue-collar and 14% for white-collar. The marriage
premiums for the model are displayed in Figures 5c and 5d for both types of wives
in both working sectors. The model allows for marriage (and divorce) to affect both
the intercept and slope, so the Þgures display the marriage premiums for workers who
work continually in the same occupation over 18 years of marriage. Most notably, the
results show that the marriage premium is practically non-existent for men married to
Type 2 wives, and perhaps might even be a little negative in the white-collar sector
(around -4%). The marriage premium with a Type 1 wife is bigger (roughly 10% in
white-collar and increasing from 9% to 13% in blue-collar over 18 years of marriage).
Clearly, these results suggest that the OLS marriage premium is biased upwards due to
25
the correlation of unobserved ability and marital status. This is particularly true since
the marriage premium appears to be close to zero for men who marry Type 2 wives,
which are typically men with higher ability and education levels. However, the results
are similar to OLS in the sense that the marriage premium is higher in the blue-collar
sector.
Table 5e and 5d demonstrate the impact of divorce on wages in both sectors.
Divorce virtually wipes out the marriage premium for men who typically earn lower
wages (those that work in blue-collar and/or marry Type 1 wives), while divorce seems
to have a short-run positive impact on high wage men (men who tend to marry Type 2
wives), but this effect dissipates over time.
Overall, the model suggests that the marriage premium is much smaller than OLS,
and is roughly 10% for men who typically have low wages (those married to Type 1
wives) and virtually non-existent for men who typically earn high wages. One possible
interpretation is that marriage changes the behavior of low wage men more than high
wage men marriage makes low wage men get serious about their careers while high
wage men are already highly motivated regardless of their martial status.
4.3 Counterfactual Experiments
The estimation of the structural parameters of the model allow for countless counterfac-
tual experiments. The following experiments were chosen to demonstrate and quantify
the importance of the interaction between marriage and career decisions, while trying
to highlight some of the factors which may have contributed heavily to many of the
trends over the last few decades in marriage and labor market outcomes.
EXPERIMENT 1: No Return to Career Decisions in the Marriage Market
This experiment is designed to see how much career decisions are governed by
their returns in the marriage market. To examine this question, the model is simulated
after turning off the possibility of marriage.10 In other words, I estimate how men would
react if there were no possibility that their actions would affect their chances of getting
10Technically, this was done by setting the probability of getting married inÞnitely small. In thesimulations, no one gets married.
26
a marriage offer, the type of marriage offer, and the chances for divorce. Figures 6a-6c
compare the differences in career decisions from the full simulated model and from the
experimental simulations. The results suggest that marriage plays a signiÞcant role
in the career decisions of men. Without the possibility to marry, men would work
less, study less, and if they do work, they would work more in blue-collar relative to
white-collar. For example, at age 25, the percent of men at home increases from
9% to 14%, the percent of men in white-collar drops from 32.8% to 21.6%, and average
years of schooling drops from 13.7 to 13.2.
These differences show that there is a signiÞcant payoff in the marriage market
to working, studying, and working in the white-collar sector. Without the returns
to these activities in the marriage market, men would make career decisions in a dis-
tinctly different manner. Furthermore, these results show that traditional estimates of
the returns to labor market decisions underestimate the true private returns without
considering their impact in the marriage market.
EXPERIMENT 2: No Exogenous Divorce
The previous experiment demonstrated the extent to which career decisions are
dominated by increasing ones chances in the marriage market. This experiment is
designed to measure the extent to which peoples decisions are determined solely by
considerations of increasing the chances of a successful marriage. To do this, I shut
off entirely the possibility of exogenous divorces in order to see whether people change
their behavior in the situation where nothing they can do will cause their wife to leave
them exogenously. (Men are still free to leave their wives in this experiment.) The
results are displayed in Figures 7a-7b.
The Þgures show that without the possibility of exogenous divorce, men work a
little more in blue-collar and go to school somewhat less. For example, at age 30 when
signiÞcant numbers of men are getting old enough to worry about divorce, the percent
of men in blue-collar increases from 42.6% to 44.7% and average years of schooling drops
from 13.92 to 13.75. These effects are not huge, and clearly the process of avoiding an
exogenous divorce is not the major factor in career decisions. But, a change of 0.17
years of schooling due only to the prospect of reducing the chances of an exogenous
27
divorce is not entirely insigniÞcant in magnitude.
EXPERIMENT 3: Increasing Transitory Shocks
Over the past few decades, the United States and other developed countries have
experienced large increases in wage inequality, and a large part of that increase is due to
increasing transitory shocks (see Gottschalk and Moffitt (1994) and Gould, Moav, and
Weinberg (2001)). This experiment is intended to see how increasing transitory shocks
affect the marriage and career decisions of young men. The experiment is conducted
by increasing the realizations of all shocks (home utility, school utility, blue-collar wage,
white-collar wage, marriage utility) by 30%, and also increasing the probability of an
exogenous divorce (given the state variables) by 30% (not 30 percentage points).
The results of this experiment capture many phenomena during the last few
decades and are presented in Figures 8a-8e. Increasing the variance of shocks causes
men to leave blue-collar work for primarily the home sector, since the amount of white-
collar workers remains roughly the same over time. At the age of 30, the percent of
men in the home sector increases from 11.0 to 15.4, while blue-collar workers fall from
42.6% to 35.6%. Interestingly, more men also seem to be going back to school in their
late 20s and early 30s, perhaps in order to deal with the increasing size of transitory
shocks and frequency of layoffs.
Also, Figure 8d shows that the average blue-collar wage increases when the vari-
ance of shocks is higher. This result is most likely due to two factors: higher shocks
produces higher wages for those who get a positive shock, and also those who received
a bad shock are now more likely to drop out of the workforce altogether, thus skim-
ming off the very low end of the blue-collar worker pool. In addition, Figure 8e shows
that the increasing variability of shocks leads to delays in marriage and lower survival
rates of marriage. For example, by the age of 25 (t=9 on the graph), the percent of
never-married men increases from 47.9 to 52.6, while the marriage survival rate at 10
years drops from 74.0% to 64.2%. These results are most likely inter-related: men are
delaying marriage in the hope of Þnding a higher quality wife in order to avoid the
higher rate of divorce.
Overall, an increasing variability of shocks produces many of the major patterns
28
in the marriage and labor markets over the past few decades: delays in marriage, higher
rates of divorce, increasing non-employment of white men, a shift away from blue-collar,
and increasing investments in education, particularly in later stages in life.
EXPERIMENT 4: Decreasing the Demand for Blue-Collar Work
This experiment is intended to describe the extent to which marriage and career
decisions have been affected by the declining demand for blue-collar work in the econ-
omy. To examine this issue, the experiment is performed by decreasing the log wage
intercept in the blue-collar sector for all types of men by 0.1. Again, the results pick
up many of the changes in the wage structure over the last few decades, as shown in
Figures 9a-9d.
Due to the declining demand for blue-collar work, workers shift in droves out of
blue-collar and into schooling and white-collar. Blue-collar wages decrease (by 11.7%
at age 30) as is expected since the demand for blue-collar work falls, but so do the wages
of white-collar workers as workers who formally worked in the blue-collar sector now
move to the bottom tail of the white-collar distribution, thus dragging down the average
white-collar wage (by 7.0% at the age of 30). There are no big changes in marriage
behavior as seen in Figure 9d, except for slightly increasing delays in marriage.
In many respects, the results from this experiment produce many of the same
patterns in the structure of the labor market that we saw from the increasing variability
of shocks in the previous experiment. The notable exception is the declining average
wage of blue-collar workers, since the previous experiment worked to increase the average
wage of those observed in the blue-collar sector. However, in contrast to the previous
experiment, the falling demand for blue-collar work had little effect on marriage market
outcomes, most likely because men started to invest more in education and white-collar
work which offset the effects of the decreasing demand for their blue-collar skill.
EXPERIMENT 5: Increasing Inequality of Wives
Since there have been large increases in male inequality, it is reasonable to ask
whether increasing variation in the quality of wives has affected the marriage and career
patterns of men. This experiment is also motivated by the strong evidence that women
29
delay their age-at-Þrst marriage in response to higher levels of male inequality (Gould
and Paserman (2003) and Loughran (2002)). To examine the behavior of men in
response to an increasing variety of women, this experiment simulates the model after
increasing the value of the higher quality type wives (Type 2 wives) by 50 percent.
The results in Figures 10a-10d show large changes in labor market behavior in
response to increasing the variance of wife quality: workers leave in droves from blue-
collar to white-collar, while education levels surge. At the age of 30, the blue-collar
sector shrinks from 42.6 to 33.8 percent, while white-collar increases from 44.1 to 53.2
percent and education levels increase by 1.45 years. In addition, Figure 10c shows
dramatic changes in marriage behavior: men delay marriage and experience more suc-
cessful marriages (higher survival rates). For example, at the age of 25, the percent of
men who are never-married increases from 47.9 to 55.5.
The reason for these changes in marriage patterns is shown in Figure 10d, where
men are now more likely to marry a Type 2 wife (the type that was increased in value).11
Overall, the increasing variation in women causes men to invest more in human capital
and white-collar experience, and delay their marriage, in order to increase the chances
of receiving a high quality offer. Therefore, this experiment also works in the direc-
tion of many of the patterns exhibited over the past few decades: more schooling, an
occupational shift towards white-collar, and delays in marriage.
EXPERIMENT 6: Increasing Divorce Costs
A natural policy instrument which affects marriage market decisions is divorce
costs. This experiment traces out the effects of doubling the costs of divorce.12 As a
result of this experiment, there were no noticeable changes in labor market decisions.
Somewhat surprisingly, men did not respond by delaying marriage to any signiÞcant
degree, however, increasing divorce costs did lower the rate of divorce (Figure 11a):
the survival rate of Þrst marriages at ten years increases from 74.0 to 78.1 percent
11In a general equilibrium setting, it would be impossible for more men to marry more Type 2 womenif the supply of Type 2 women is Þxed. Therefore, we should think of these experiments as being theeffect of manipulating the marginal person.12Technically, the current-period costs of divorce were doubled and the terminal value of divorce was
doubled.
30
which is a 15.8% reduction in the divorce rate at ten years of marriage.
EXPERIMENT 7: No Possibility for Re-Marriage
This experiment examines the extent to which mens decisions are inßuenced by
the option of re-marriage. Consequently, the probability of receiving a marriage offer
in this experiment is set to zero if you are or ever have been married. Although
not shown, there were no noticeable changes in career decisions in response to this
experiment, except for a small increase in education levels (about 0.07 of a year at
age 30), as seen in Figure 12a. However, Figure 12b shows that the percent of Þrst
marriages which survive at least 10 years increases from 74.0 to 77.5. This increase
represents a 13.5% reduction in the divorce rate at ten years of marriage. One possible
inference from this result is that a declining stigma of being divorced, which affects
the chances of re-marrying, may have contributed in a signiÞcant way to the increasing
incidence of divorce over time.
5 Conclusion
The existing literature on human capital and occupational decisions assume that indi-
viduals maximize their returns to various labor market opportunities, but this literature
ignores the evolutionary instinct in men to achieve material success in order to attract
female partners. This paper demonstrates that this instinct is alive and well even in
todays modern economy. If there were no returns to career outcomes in the marriage
market, the results suggest that men would work less, study less, and if they did work,
they would work more in the blue-collar sector than the white-collar sector. Over-
looking these factors, as is done in the large literature estimating the returns to labor
market decisions, underestimates the true private returns to career decisions.
Beyond showing the importance of the interaction between career choices and
marriage outcomes and vice versa, this paper also sheds light on many of the labor
market and marriage patterns in the data over the last few decades. In particular, the
results indicate that the increasing variation of transitory shocks, coupled with the de-
31
clining demand for blue-collar wages, led to increasing non-employment, more schooling,
an occupational shift from blue-collar to white-collar, declining wages in both sectors,
delays in marriage, and increasing divorce rates. In addition, increasing inequality of
females may also have contributed to many of these trends. In particular, increasing
inequality of women was shown to cause men to delay marriage while searching and
competing more heavily for high quality women, by increasing their education levels
and by working more in white-collar as opposed to blue-collar.
In addition, the parameter estimates indicate that traditional OLS estimates of
the marriage premium are biased upwards and stem largely from unobserved hetero-
geneity. After controlling for unobserved heterogeneity and endogenous choices, the
marriage premium is estimated to be zero for men who typically earn high wages and
roughly 11% for those men who earn lower wages. One possible interpretation for these
results is that marriage makes lower wage men more motivated and serious about their
responsibilities, while marriage has little effect on high wage men since they are already
highly motivated.
All of these results point to the importance of considering the interactive nature
of the joint career and marriage decisions of men, a process which has been overlooked
thus far in the existing empirical literature.
32
6 Bibliography1. Becker, Gary S. (1991). A Treatise on the Family, Enlarged edition. Cambridge,MA: Harvard University Press.
2. Becker, Gary S.; Landes, Elizabeth M.; and Robert T. Michel (1977).An Eco-nomic Analysis of Marital Instability. Journal of Political Economy 85: 1141-87.
3. Bellman, R. (1957), Dynamic Programming, Princeton, New Jersey: PrincetonUniversity Press.
4. Bergstrom, Theodore; and Mark Bagnoli (1993). Courtship as a Waiting GameJournal of Political Economy, 101: 185-202.
5. Brien, Michael J.; Lillard, Lee A.; and Stern, Steven, (2002) Cohabitation, Mar-riage, and Divorce in a Model of Match Quality, forthcoming in the InternationalEconomic Review.
6. Chiappori, P., Fortin, B, and G. Lacroix (2002) Marriage Market, Divorce Leg-islation, and Household Labor Supply, Journal of Political Economy 110 (1):37-72.
7. Cohen, Yinon and Haberfeld, Yitchak. (1991) Why Do Married Men EarnMore than Unmarried Men? Social Science Research 20, 29-44.
8. Eckstein, Z. and Wolpin, K.I. (1989), Dynamic Labour Force Participation ofMarriedWomen and EndogenousWork Experience, Review of Economic Studies,56, 375-390.
9. Fernandez, Raquel; Guner, Nezih; and Knowles, John. (2001) Love and Money:ATheoretical and Empirical Analysis of Household Sorting and Inequality, CEPRWorking Paper 3040.
10. Gottschalk, P. and Moffitt, R. (1994) The Growth in Earnings Instability in theU.S. Labor Market, Brookings Papers on Economic Activity 2, 217-254.
11. Gould, Eric D. (2002) Rising Wage Inequality, Comparative Advantage, andthe Growing Importance of General Skills in the United States Journal of LaborEconomics 20(1): 105-47
12. Gould, Eric D., and M. Daniele Paserman (2003) Waiting for Mr. Right: RisingInequality and Declining Marriage Rates, Journal of Urban Economics, 53: 257-281.
13. Gould, Eric D.; Omer Moav, and Bruce A. Weinberg (2001) Precautionary De-mand for Education, Inquality, and Technological Progress, Journal of EconomicGrowth 6: 285-315.
14. Gray, Jeffrey S. (1997), The Fall in Mens Return to Marriage: DecliningProductivity Effects or Changing Selection? Journal of Human Resources 32(3): 481-504.
15. Heckman, J. and MaCurdy, T. (1980), A Life Cycle Model of Female LabourSupply, Review of Economic Studies, 47, 47-74.
33
16. Heckman, J., and B. Singer (1984), A Method for Minimizing the Impact ofDistributional Assumptions in Econometric Models for Duration Data, Econo-metrica, 2, 271-320.
17. Heckman, James J., (1981) The Incidental Parameters Problem and the Problemof Initial Conditions in Estimating a Discrete Time-Discrete Data Stochastic Pro-cess and Some Monte Carlo Evidence in C. F. Manksi and D. McFadden, eds.,Structural Analysis of Discrete Data with Econometric Applications (Cambridge,MA: MIT Press, 1981).
18. Hersch, Joni; and Leslie Stratton (2000). Household Specialization and the MaleMarriage Wage Premium. Industrial and Labor Relations Review 54(1): 78-94.
19. Hotz, V.J., and Miller, R.A. (1988), An Empirical Analysis of Life Cycle Fertilityand Female Labor Supply, Econometrica, 56, 91-118.
20. Hotz, V.J.; Miller, R.A.; Sanders, Seth; and Smith, Jeffrey (1994) A SimulationEstimator for Dynamic Models of Discrete Choice, Review of Economic Studies61(2): 265-289.
21. Keane, Michael.P., and Kenneth I. Wolpin (1994), The Solution and Estimationof Discrete Choice Dynamic Programming Models by Simulation and Interpola-tion: Monte Carlo Evidence, Review of Economics and Statistics, 17, 648-672.
22. Keane, Michael P.; and Kenneth I. Wolpin (1997). The Career Decisions ofYoung Men. Journal of Political Economy 105: 473-522.
23. Keane, Michael P.; and Kenneth I. Wolpin (2001). The Effect of BorrowingConstraints on Educational Attainment International Economic Review 42(4):1051-1103.
24. Korenman, Sanders; and David Neumark (1991). Does Marriage Really MakeMen More Productive? Journal of Human Resources 26(2): 282-307.
25. Loughran, D.S. (2002). The Effect of Male Wage Inequality on Female Age-At-First Marriage, Review of Economics and Statistics 84: 237-251.
26. Mortensen, D. (1988) Matching: Finding a Partner for Life or Otherwise, Amer-ican Journal of Sociology, 94, S215-S240.
27. Rust, J. (1987). Optimal Replacement of GMC Bus Engines: An EmpiricalModel of Howard Zurcher, Econometrica, 55(5), 999-1033.
28. Seitz, Shannon N., (2002), Accounting for Racial Differences in Marriage andEmployment, working paper, Queens University.
29. Taber, Christopher R. (2001). The Rising College Premium in the Eighties:Return to College or Return to Unobserved Ability? Review of Economic Studies68 (3): 665-691.
30. van der Klaauw, Wilbert (1996). Female Labour Supply and Marital StatusDecisions: A Life-Cycle Model, Review of Economic Studies, 63, 199-235.
31. Weiss, Yoram (1997). The Formation and Dissolution of Families: Why Marry?Who Marries Whom? And What Happens upon Marriage and Divorce? inHandbook of Population and Family Economics, edited by Robert Rosenzweigand Oded Stark. Amsterdam: Elsevier Science, Volume 1A: 81-124.
34
32. Weiss, Yoram; and Robert J. Willis (1997). Match Quality, New Information,and Marital Dissolution Journal of Labor Economics, 15: S293-S329.
33. Wong, Linda. Structural Estimation of Marriage Models SSRNWorking Paper,University of Chicago and Binghampton University.
35
Table 1: Marriage Choices Over TimeAge Never
Married(Single)
FirstMarriage
FirstDivorce(Single)
SecondMarriage
SecondDivorce(Single)
ThirdMarriage
SampleSize
16 1.0 0.00 0.00 0.00 0.00 0.00 175617 0.99 0.01 0.00 0.00 0.00 0.00 175618 0.96 0.04 0.00 0.00 0.00 0.00 175419 0.91 0.09 0.00 0.00 0.00 0.00 175320 0.85 0.15 0.01 0.00 0.00 0.00 175121 0.78 0.21 0.01 0.00 0.00 0.00 174522 0.70 0.27 0.02 0.01 0.00 0.00 174123 0.63 0.33 0.03 0.01 0.00 0.00 173424 0.56 0.38 0.04 0.02 0.00 0.00 173225 0.48 0.44 0.05 0.03 0.00 0.00 173426 0.43 0.47 0.06 0.03 0.00 0.00 173227 0.38 0.50 0.07 0.04 0.01 0.00 173128 0.34 0.53 0.08 0.05 0.01 0.00 172629 0.31 0.54 0.08 0.06 0.01 0.01 172530 0.28 0.55 0.09 0.07 0.01 0.01 171131 0.26 0.56 0.09 0.08 0.01 0.01 170332 0.23 0.57 0.09 0.09 0.01 0.01 168433 0.21 0.58 0.09 0.09 0.01 0.01 168434 0.19 0.59 0.10 0.09 0.02 0.01 144035 0.17 0.60 0.10 0.09 0.02 0.02 121336 0.16 0.60 0.09 0.10 0.03 0.02 94937 0.14 0.60 0.10 0.11 0.03 0.02 73638 0.14 0.59 0.10 0.11 0.03 0.03 51939 0.11 0.64 0.10 0.11 0.04 0.02 332
The numbers represent the percent of the sample of men in each marital status. The suppressedcategory is for those who are divorced three times and single, which is never above one percent of thesample.
Table 2: Career Choices Over Time
Age Home SchoolBlue
CollarWhiteCollar
SampleSize
16 0.06 0.91 0.03 0.00 92317 0.08 0.86 0.06 0.00 114518 0.12 0.59 0.24 0.04 137019 0.13 0.40 0.38 0.09 157920 0.14 0.33 0.42 0.11 176121 0.13 0.28 0.45 0.15 176222 0.12 0.17 0.48 0.22 177823 0.12 0.11 0.49 0.28 175024 0.10 0.07 0.49 0.34 173325 0.09 0.06 0.48 0.37 171126 0.08 0.04 0.49 0.39 169227 0.08 0.03 0.49 0.41 168228 0.08 0.02 0.48 0.42 167629 0.07 0.03 0.47 0.43 146830 0.08 0.02 0.48 0.42 146931 0.08 0.03 0.45 0.44 122132 0.08 0.02 0.44 0.45 124333 0.08 0.03 0.43 0.46 97434 0.08 0.03 0.43 0.46 83835 0.08 0.03 0.42 0.47 57136 0.06 0.03 0.42 0.49 38937 0.11 0.03 0.40 0.47 36138 0.09 0.04 0.42 0.46 17139 0.08 0.03 0.38 0.51 170
Each number represents the percent of the sample of men in each of the mutually exclusive andexhaustive career categories.
Table 3: Log Wage Regressions for NLSY79 Sample of Workers in 1992
All Workers Blue-Collar Workers White Collar WorkersIntercept 4.35 (0.35) 4.07 (0.63) 4.29 (0.46)Education 0.06 (0.01) 0.04 (0.02) 0.07 (0.01)Experience 0.05 (0.03) 0.09 (0.06) 0.03 (0.05)Experience Squared -0.00 (0.00) -0.00 (0.00) -0.00 (0.00)South -0.05 (0.03) -0.06 (0.04) -0.04 (0.04)SMSA 0.15 (0.03) 0.15 (0.04) 0.15 (0.05)AFQT 0.07 (0.02) 0.08 (0.02) 0.07 (0.03)# Children Household 0.03 (0.01) 0.01 (0.02) 0.06 (0.03)Both Parents at 14 0.06 (0.04) 0.08 (0.05) 0.03 (0.06)Mother HS Grad -0.01 (0.04) -0.01 (0.04) 0.02 (0.06)Mother College Grad 0.08 (0.06) 0.05 (0.09) 0.09 (0.08)Father HS Graduate 0.08 (0.03) 0.09 (0.04) 0.09 (0.06)Father College Grad 0.05 (0.05) 0.01 (0.07) 0.09 (0.07)Blue-Collar -0.17 (0.03)Married 0.18 (0.03) 0.22 (0.05) 0.14 (0.05)Divorced 0.04 (0.05) 0.05 (0.06) 0.05 (0.10)
# Observations 1410 711 699R-square 0.26 0.15 0.16Standard errors in parentheses. Wages are defined as weekly wages. Experience is equal to age minuseducation minus six. High school graduates for mothers and fathers include those who started collegebut did not finish. SMSA is a dummy variable for living in a SMSA.
Table 4: Probit Regressions on Marital Status for NLSY79 Sample in 1992
Probability of Being Married Probability of Being Divorced
Intercept -1.59 (0.51) -4.56 (0.68) -1.01 (0.77) 0.94 (0.99)Home -0.76 (0.13) 0.87 (0.18)School -0.27 (0.22) -0.10 (0.50)Blue-Collar -0.33 (0.08) -0.26 (0.09) 0.41 (0.12) 0.39 (0.12)Education -0.01 (0.02) -0.04 (0.02) -0.12 (0.03) -0.09 (0.03)Age 0.08 (0.01) 0.08 (0.02) 0.05 (0.02) 0.05 (0.02)South 0.08 (0.07) 0.11 (0.08) 0.21 (0.09) 0.19 (0.11)SMSA -0.18 (0.07) -0.27 (0.08) 0.13 (0.10) 0.12 (0,11)AFQT 0.05 (0.04) -0.00 (0.00) 0.02 (0.06) 0.06 (0.07)Both Parents at 14 0.24 (0.08) 0.30 (0.10) -0.19 (0.12) -0.25 (0.14)Mother HS Grad 0.05 (0.09) 0.03 (0.10) -0.13 (0.12) 0.09 (0.13)Mother College Grad -0.27 (0.14) -0.24 (0.16) 0.31 (0.22) 0.31 (0.24)Father HS Graduate -0.16 (0.08) -0.20 (0.09) 0.22 (0.12) 0.27 (0.13)Father College Grad -0.14 (0.11) -0.12 (0.12) 0.02 (0.18) 0.02 (0.20)# Children Born -0.17 (0.04) -0.15 (0.05)Log Weekly Wage 0.55 (0.07) -0.40 (0.11)
# Observations 1689 1410 1223 1044Sample Full Sample Workers Only Married and/or
DivorcedSample
Only Workerswho are Marriedand/or Divorced
Standard errors in parentheses.
Table 5: Actual and Model Predicted Marriage Transition Matrix
Status in Period t+1
Status inPeriod t
NeverMarried(single)
FirstMarriage
FirstDivorce(single)
SecondMarriage
SecondDivorce(single)
ThirdMarriage
ThirdDivorce(single)
RowSampleSize inNLSY
NeverMarried(single)
0.92
0.93
0.08
0.07
18,429
FirstMarriage
0.97
0.97
0.03
0.03
0.0
0.0
12,729
FirstDivorce(single)
0.86
0.87
0.14
0.13
1,680
SecondMarriage
0.94
0.88
0.05
0.09
0.0
0.0
1,284
SecondDivorce(single)
0.85
0.81
0.15
0.19
204
ThirdMarriage
0.93
0.85
0.07
0.15
122
The number in the upper left side of each box represents the actual transitions in the NLSY data(percent of persons in the row’s marriage category which move to the category in the column), whilethe numbers in the lower right hand side of each box represent the model predictions. The empty boxesrepresent transitions which defy a logical order of marriage situations, so they are by definition empty.Persons who transition into the “third marriage” are not allowed to transition out of this category in theanalysis.
Table 6: Actual and Model Predicted Career Choices Transition Matrix
Choice in Period t+1
Choice inPeriod t
School BlueCollar
WhiteCollar
Home RowSample Size
in NLSY
School0.64
0.60
0.14
0.15
0.14
0.16
0.09
0.09
5,143
Blue Collar0.02
0.02
0.80
0.79
0.12
0.09
0.06
0.10
9,864
WhiteCollar
0.04
0.04
0.16
0.10
0.78
0.80
0.03
0.06
6,414
Home0.11
0.09
0.31
0.38
0.11
0.26
0.47
0.28
2,272
The number in the upper left side of each box represents the actual transitions in the NLSY data(percent of persons in the row’s career category which move to the category in the column), while thenumbers in the lower right hand side of each box represent the model predictions.
Perc
ent in
Sec
tor
Figure 1a: Model Fit Career Choices (Home and Blue Collar)age
Actual Home Sector Model Home Sector Actual Blue-Collar Sector Model Blue-Collar Sector
16 18 20 22 24 26 28 30 32 34 36 38
0
.2
.4
.6
.8
1
Perc
ent i
n Se
ctor
Figure 1b: Model Fit of Career Choices (School and White Collar)age
Actual School Sector Model School Sector Actual White-Collar Sector Model White-Collar Sector
16 18 20 22 24 26 28 30 32 34 36 38
0
.2
.4
.6
.8
1
Mea
n Lo
g W
age
Figure 2: Model Fit of White and Blue-Collar Wagesage
Actual Blue-Collar Wage Model Blue-Collar Wage Actual White-Collar Wage Model White-Collar Wage
16 18 20 22 24 26 28 30 32 34 36 38
4.5
4.7
4.9
5.1
5.3
5.5
5.7
5.9
6.1
6.3
6.5
Perc
ent o
f Pop
ulat
ion
Figure 3a: Model Fit of Marriage Statusage
Actual Never Married Model Never Married Actual First Marriage Model First Marriage
16 18 20 22 24 26 28 30 32 34 36 38
0
.2
.4
.6
.8
1
Perc
ent o
f Pop
ulat
ion
Figure 3b: Model Fit of Marriage Statusage
Actual First Divorce Model First Divorce Actual Second Marriage Model Second Marriage
16 18 20 22 24 26 28 30 32 34 36 38
0
.05
.1
.15
.2
Perc
ent o
f Pop
ulat
ion
Figure 3c: Model Fit of Marriage Statusage
Actual Second Divorce Model Second Divorce Actual Third Marriage Model Third Marriage
16 18 20 22 24 26 28 30 32 34 36 38
0
.01
.02
.03
.04
.05
Sur
viva
l Rat
e
Figure 3d: Fit of First Marriage Survival Ratetime period
Actual Survival Rate Model Marriage Survival Rate
0 2 4 6 8 10 12 14
.5
.6
.7
.8
.9
1
Figure 4a: Model Fit of Percent Never-Married by AFQT
Actual Percent Never-Married Model Percent Never-MarriedBottom 0 to 20th AFQT Percentile
0.2.4.6.81
Bottom 20th to 40th AFQT Percent Bottom 40th to 60th AFQT Percent
0 2 4 6 8 10 12 14Top 60th to 80th AFQT Percentile
0 2 4 6 8 10 12 140.2.4.6.81
Top 80th to 100 AFQT Percentile
0 2 4 6 8 10 12 14
Figure 4b: Model Fit of First Marriage Survival Rates by AFQT
Actual Survival Rate Model Survival RateBottom 0 to 20th AFQT Percentile
.5
.7
.9
Bottom 20th to 40th AFQT Percent Bottom 40th to 60th AFQT Percent
0 2 4 6 8 10 12 14Top 60th to 80th AFQT Percentile
0 2 4 6 8 10 12 14.5
.7
.9
Top 80th to 100 AFQT Percentile
0 2 4 6 8 10 12 14
Figure 4c: Model Fit of Career Choices by AFQT
Actual Home Sector Model Home Sector Actual Blue-Collar Sector Model Blue-Collar Sector
Bottom 0 to 20th AFQT Percentile
0.2.4.6.81
Bottom 20th to 40th AFQT Percent Bottom 40th to 60th AFQT Percent
16 20 24 28 32 36Top 60th to 80th AFQT Percentile
16 20 24 28 32 360.2.4.6.81
Top 80th to 100 AFQT Percentile
16 20 24 28 32 36
Figure 4d: Model Fit of Career Choices by AFQT
Actual School Sector Model School Sector Actual White-Collar Sector Model White-Collar Sector
Bottom 0 to 20th AFQT Percentile
0.2.4.6.81
Bottom 20th to 40th AFQT Percent Bottom 40th to 60th AFQT Percent
16 20 24 28 32 36Top 60th to 80th AFQT Percentile
16 20 24 28 32 360.2.4.6.81
Top 80th to 100 AFQT Percentile
16 20 24 28 32 36
Mea
n Lo
g W
age
Figure 4e: Model Fit of Wages by AFQT
Actual Blue-Collar Wage Model Blue-Collar Wage Actual White-Collar Wage Model White-Collar Wage
Bottom 0 to 20th AFQT Percentile
4.54.74.95.15.35.55.75.96.16.36.5
Bottom 20th to 40th AFQT Percent Bottom 40th to 60th AFQT Percent
161820222426283032343638Top 60th to 80th AFQT Percentile
1618202224262830323436384.54.74.95.15.35.55.75.96.16.36.5
Top 80th to 100 AFQT Percentile
161820222426283032343638
Perc
ent o
f Men
Figure 5a: Unobserved Wife Typesage
Model Percent Single Model Married to Wife Type 1 Model Married to Wife Type 2
16 18 20 22 24 26 28 30 32 34 36 38
0
.2
.4
.6
.8
1
Pe
rce
nt o
f Men
Figure 5b: Unobserved Wife Types by AFQTage
Model Percent Single Model Married to Wife Type 1 Model Married to Wife Type 2
Bottom 0 to 20th AFQT Percentile
0.2.4.6.81
Bottom 20th to 40th AFQT Percent Bottom 40th to 60th AFQT Percent
161820222426283032343638Top 60th to 80th AFQT Percentile
1618202224262830323436380.2.4.6.81
Top 80th to 100 AFQT Percentile
161820222426283032343638
Log
Wag
es R
elat
ive
to U
nmar
ried
Men
Figure 5c: Marriage Premium with Type 1 Wifetime period
Blue-Collar Marriage Premium White-Collar Marriage Premium
0 2 4 6 8 10 12 14 16 18
.09
.1
.11
.12
.13
Log
Wag
es R
elat
ive
to U
nmar
ried
Men
Figure 5d: Marriage Premium with Type 2 Wifetime period
Blue-Collar Marriage Premium White-Collar Marriage Premium
0 2 4 6 8 10 12 14 16 18
-.04
-.02
0
.02
.04
Log
Wag
es R
elat
ive
to U
nmar
ried
Men
Figure 5e: Marriage and Divorce Premiums with Type 1 Wifetime period
Blue-Collar Marriage Premium White-Collar Marriage Premium Divorcing at t=5, Blue-Collar Divorcing at t=5, White-Collar
0 2 4 6 8 10 12 14 16 18
-.05
0
.05
.1
.15Lo
g W
ages
Rel
ativ
e to
Unm
arrie
d M
en
Figure 5f: Marriage and Divorce Premiums with Type 2 Wifetime period
Blue-Collar Marriage Premium White-Collar Marriage Premium Divorcing at t=5, Blue-Collar Divorcing at t=5, White-Collar
0 2 4 6 8 10 12 14 16 18
-.05
0
.05
Per
cent
in S
ecto
r
Figure 6a: Experiment 1 Career Choices (Home and Blue Collar)age
Model Home Sector EXP Home Sector Model Blue-Collar Sector EXP Blue-Collar Sector
16 18 20 22 24 26 28 30 32 34 36 38
0
.2
.4
.6
.8
1
Perc
ent in
Sec
tor
Figure 6b: Experiment 1 Career Choices (School and White Collar)age
Model School Sector EXP School Sector Model White-Collar Sector EXP White-Collar Sector
16 18 20 22 24 26 28 30 32 34 36 38
0
.2
.4
.6
.8
1
Mea
n Ye
ars
of E
duca
tion
Figure 6c: Experiment 1 Years of Educationage
Model Mean Years of Education EXP Mean Years of Education
16 18 20 22 24 26 28 30 32 34 36 38
11
12
13
14
Perc
ent i
n Se
ctor
Figure 7a: Experiment 2 Career Choices (Home and Blue Collar)age
Model Home Sector EXP Home Sector Model Blue-Collar Sector EXP Blue-Collar Sector
16 18 20 22 24 26 28 30 32 34 36 38
0
.2
.4
.6
.8
1
Perc
ent i
n Se
ctor
Figure 7b: Experiment 2 Career Choices (School and White Collar)age
Model School Sector EXP School Sector Model White-Collar Sector EXP White-Collar Sector
16 18 20 22 24 26 28 30 32 34 36 38
0
.2
.4
.6
.8
1
Mea
n Y
ears
of E
duca
tion
Figure 7c: Experiment 2 Years of Educationage
Model Mean Years of Education EXP Mean Years of Education
16 18 20 22 24 26 28 30 32 34 36 38
11
12
13
14
Per
cent
in S
ecto
r
Figure 8a: Experiment 3 Career Choices (Home and Blue Collar)age
Model Home Sector EXP Home Sector Model Blue-Collar Sector EXP Blue-Collar Sector
16 18 20 22 24 26 28 30 32 34 36 38
0
.2
.4
.6
.8
1
Per
cent
in S
ecto
r
Figure 8b: Experiment 3 Career Choices (School and White Collar)age
Model School Sector EXP School Sector Model White-Collar Sector EXP White-Collar Sector
16 18 20 22 24 26 28 30 32 34 36 38
0
.2
.4
.6
.8
1
Mea
n Y
ears
of E
duca
tion
Figure 8c: Experiment 3 Years of Educationage
Model Mean Years of Education EXP Mean Years of Education
16 18 20 22 24 26 28 30 32 34 36 38
11
12
13
14
15
Mea
n Lo
g W
age
Figure 8d: Experiment 3 White and Blue-Collar Wagesage
Model Blue-Collar Wage EXP Blue-Collar Wage Model White-Collar Wage EXP White-Collar Wage
16 18 20 22 24 26 28 30 32 34 36 38
4.5
4.7
4.9
5.1
5.3
5.5
5.7
5.9
6.1
6.3
6.5
Perc
ent
Figure 8e: Experiment 3 Single and First Marriage Survival Ratestime period
Model Percent Never-Married EXP Percent Never-Married Model Marriage Survival Rate EXP Marriage Survival Rate
0 2 4 6 8 10 12 14
.5
.6
.7
.8
.9
1
Perc
ent i
n Se
ctor
Figure 9a: Experiment 4 Career Choices (Home and Blue Collar)age
Model Home Sector EXP Home Sector Model Blue-Collar Sector EXP Blue-Collar Sector
16 18 20 22 24 26 28 30 32 34 36 38
0
.2
.4
.6
.8
1
Perc
ent i
n Se
ctor
Figure 9b: Experiment 4 Career Choices (School and White Collar)age
Model School Sector EXP School Sector Model White-Collar Sector EXP White-Collar Sector
16 18 20 22 24 26 28 30 32 34 36 38
0
.2
.4
.6
.8
1
Mea
n Lo
g W
age
Figure 9c: Experiment 4 White and Blue-Collar Wagesage
Model Blue-Collar Wage EXP Blue-Collar Wage Model White-Collar Wage EXP White-Collar Wage
16 18 20 22 24 26 28 30 32 34 36 38
4.5
4.7
4.9
5.1
5.3
5.5
5.7
5.9
6.1
6.3
6.5
Perc
ent
Figure 9d: Experiment 4 Single and First Marriage Survival Ratestime period
Model Percent Never-Married EXP Percent Never-Married Model Marriage Survival Rate EXP Marriage Survival Rate
0 2 4 6 8 10 12 14
.5
.6
.7
.8
.9
1
Perc
ent in
Sec
tor
Figure 10a: Experiment 5 Career Choices (Home and Blue Collar)age
Model Home Sector EXP Home Sector Model Blue-Collar Sector EXP Blue-Collar Sector
16 18 20 22 24 26 28 30 32 34 36 38
0
.2
.4
.6
.8
1
Perc
ent i
n Se
ctor
Figure 10b: Experiment 5 Career Choices (School and White Collar)age
Model School Sector EXP School Sector Model White-Collar Sector EXP White-Collar Sector
16 18 20 22 24 26 28 30 32 34 36 38
0
.2
.4
.6
.8
1
Perc
ent
Figure 10c: Experiment 5 Single and First Marriage Survival Ratestime period
Model Percent Never-Married EXP Percent Never-Married Model Marriage Survival Rate EXP Marriage Survival Rate
0 2 4 6 8 10 12 14
.5
.6
.7
.8
.9
1Pe
rcen
t of M
en
Figure 10d: Experiment 5 Unobserved Wife Typesage
Model Married to Wife Type 1 EXP Married to Wife Type 1 Model Married to Wife Type 2 EXP Married to Wife Type 2
16 18 20 22 24 26 28 30 32 34 36 38
0
.2
.4
.6
Perc
ent
Figure 11a: Experiment 6 Single and First Marriage Survival Ratestime period
Model Percent Never-Married EXP Percent Never-Married Model Marriage Survival Rate EXP Marriage Survival Rate
0 2 4 6 8 10 12 14
.5
.6
.7
.8
.9
1
Perc
ent o
f Men
Figure 11b: Experiment 6 Unobserved Wife Typesage
Model Married to Wife Type 1 EXP Married to Wife Type 1 Model Married to Wife Type 2 EXP Married to Wife Type 2
16 18 20 22 24 26 28 30 32 34 36 38
0
.2
.4
.6
Mea
n Ye
ars
of E
duca
tion
Figure 12a: Experiment 7 Years of Educationage
Model Mean Years of Education EXP Mean Years of Education
16 18 20 22 24 26 28 30 32 34 36 38
11
12
13
14
Perc
ent
Figure 12b: Experiment 7 Single and First Marriage Survival Ratestime period
Model Percent Never-Married EXP Percent Never-Married Model Marriage Survival Rate EXP Marriage Survival Rate
0 2 4 6 8 10 12 14
.5
.6
.7
.8
.9
1
Appendix: Marriage and Career Model EstimatesLog Wage Functions
Blue Collar White Collar
Type 1 intercept 4.79 (0.0023) 4.54 (0.0032)Type 1 experience 0.06 (0.0003) 0.06 (0.0005)Type 1 expersquare/100
-0.20 (0.0027) -0.26 (0.0027)
Type 2 intercept 5.06 (0.0027) 4.67 (0.0026)Type 2 experience 0.05 (0.0002) 0.05 (0.0004)Type 2 expersquare/100
-0.18 (0.0016) -0.19 (0.0036)
Type 3 intercept 5.01 (0.0025) 4.77 (0.0024)Type 3 experience 0.05 (0.0002) 0.07 (0.0003)Type 3 expersquare/100
-0.05 (0.0007) -0.14 (0.0014)
HS Graduate 0.11 (0.0012) 0.17 (0.0016)College Graduate 0.08 (0.0012) 0.23 (0.0012)Education 0.03 (0.0003) 0.06 (0.0002)
Married to Type 1 Wife 0.09 (0.0011) 0.10 (0.0012)Married to Type 2 Wife -0.005 (0.0057) -0.04 (0.0013)Ever Divorced 0.03 (0.0008) 0.03 (0.0006)Ever Married*experience
0.002 (0.0001) -0.0005 (0.00003)
Divorced*experience -0.006 (0.0001) -0.004 (0.0003)
AFQT 0.03 (0.0009) 0.18 (0.0013)AFQT*experience 0.01 (0.0001) 0.003 (0.0001)AFQT*experiencesquared/100
0.02 (0.0004) -0.01 (0.0006)
Under 18 -0.18 (0.0062) -0.58 (0.0346)Experience in the otheroccupation
0.04 (0.0002) 0.06 (0.0002)
Experience squared inother occupation
-0.14 (0.0019) -0.15 (0.0014)
Worked in sameoccupation in previousperiod
0.15 (0.0012) 0.19 (0.0016)
Probability Functions for Types of Men(multivariate logit)
Type 2 Type 3Intercept -0.99 (0.0052) -1.59 (0.0052)AFQT -0.20 (0.0891) 0.04 (0.0694)Lived with Both Parentsat age 14
0.34 (0.1061) 1.15 (0.0067)
Mother HS Grad 0.31 (0.0087) 0.60 (0.0060)Mother College Grad 0.36 (0.2597) 0.91 (0.0143)Father HS Grad 0.35 (0.1095) 0.44 (0.0066)Father College Grad -0.04 (0.2129) 0.73 (0.0071)
Estimated Percent ofMen of Each Type
24% 43%
Marriage Offer Functions(logit functions)
Type 1 Wife Type 2 WifeIntercept -5.88 (0.0037) -9.50 (0.0095)Education -0.04 (0.0003) 0.45 (0.0019)Blue Collar Exper 0.11 (0.0006) 0.17 (0.0022)White Collar Exper 0.24 (0.0009) 0.21 (0.0022)Age -0.01 (0.0001) 0.002 (0.0002)Age Squared -0.002 (0.0000) -0.01 (0.00004)HS Grad 0.31 (0.0020) 0.62 (0.0053)College Grad 0.55 (0.0068) 2.24 (0.0117)Currently in HighSchool
-0.64 (0.0388) 0.17 (0.3076)
Currently in Post HighSchool Studies
-0.67 (0.0093) -1.19 (0.0158)
Working in Blue Collar 1.41 (0.0040) 0.94 (0.0106)Working in WhiteCollar
1.24 (0.0050) 1.14 (0.0105)
Currently in firstmarriage
-10.75 (1977.94) -4.95 (0.0356)
Ever Divorced -2.19 (0.0042) -3.74 (0.0331)Divorced and Single 0.76 (0.0043) 0.89 (1.0173)Type 2 Male 3.15 (0.0045) 2.00 (0.0162)Type 3 Male 1.89 (0.0079) 2.74 (0.0092)AFQT -0.28 (0.0019) -0.37 (0.0022)
Exogenous Divorce Function(logit function)
Currently Home 0.51 (0.0007)Currently in School -0.38 (0.0194)Currently in Blue Collar -0.47 (0.0187)Currently in WhiteCollar
-0.47 (0.0135)
Education -0.13 (0.0024)Type 1 Male -3.78 (0.0168)Type 2 Male -2.25 (0.0036)Type 3 Male -5.35 (0.0067)AFQT -0.29 (0.0066)
Current Period Non-Pecuniary UtilitiesHome Sector(log utility)
Type 1 Male intercept 5.09 (0.0052)Type 2 Male intercept 5.07 (0.0017)Type 3 Male intercept 4.89 (0.0084)Age 0.03 (0.0002)Age squared/100 -0.12 (0.0010)AFQT -0.12 (0.0013)
School SectorIntercept -15.32 (0.3863)Under 18 205.84 (5.4229)
MarriedType 1 Wife 158.09 (0.6224)Type 2 Wife 277.22 (1.3446)
Current Period CostsNet Tuition Costs
High School 506.25 (8.1204)College -112.78 (0.8419)
Post-College 204.07 (2.2276)Entry Costs into School
Not HS Grad -257.94 (5.6719)HS Grad -296.56 (1.4645)
Entry Costs into OccupationsBlue Collar White Collar
Intercept -118.34 (1.0178) -141.21 (1.0758)Age -8.34 (0.1084) -1.34 (0.0508)
Divorce CostsIntercept -0.88 (162.057)Divorced at least twice -50.37 (0.8406)Divorced three times -48.72 (42.8735)
Terminal Value Emax FunctionWorking in Blue Collar 221.97 (6.2864)Working in WhiteCollar
162.46 (3.8943)
Education 37.76 (1.0555)Blue Collar Exper 69.01 (0.6840)White Collar Exper 30.92 (0.4667)Education*AFQT 129.02 (1.9054)Blue CollarExper*AFQT
12.82 (0.4414)
White CollarExper*AFQT
25.42 (0.5629)
Divorced at least once -1154.13 (17.2341)Divorced at least twice -552.16 (12.8990)Divorced three times -58.00 (10673.03)
Cholesky Decomposition Matrix of Shocks
Log BlueCollarWage
Log WhiteCollarWage
Log HomeUtility
SchoolUtility
MarriageUtility
Log BlueCollar Wage
0.51(0.0015)
Log WhiteCollar Wage
-0.06(0.0009)
0.51(0.0016)
Log HomeUtility
-0.04(0.0008)
0.004(0.0004)
0.92(0.0019)
SchoolUtility
122.64(0.7963)
15.30(0.7705)
-0.74(0.8103)
305.89(0.7588)
MarriageUtility
453.51(2.4946)
149.75(5.3299)
-0.60(2.3270)
216.15(5.6826)
-189.13(2.7051)
Standard Deviation of Shocks(implied by Cholesky Decomposition of shocks)0.51 0.51 0.92 329.91 557.30
Correlation of ShocksLog BlueCollar Wage
1.0
Log WhiteCollar Wage
-0.12 1.0
Log HomeUtility
-0.04 0.01 1.0
SchoolUtility
0.37 -0.00 -0.02 1.0
MarriageUtility
0.81 0.17 -0.04 0.00 1.0
Initial Conditions of Parental BackgroundLived with Both Parents at age 14
(logit function)Intercept 1.62 (0.0631)AFQT 0.33 (0.0714)
Mother’s Education(multivariate logit)
High School Graduate College GraduateIntercept 0.99 (0.1356) -0.99 (0.2267)AFQT 0.85 (0.0741) 1.71 (0.1186)Lived with BothParents at 14
0.29 (0.1463) 0.25 (0.2420)
Father’s Education(multivariate logit)
High School Graduate College GraduateIntercept -0.78 (0.1665) -3.27 (0.3409)AFQT 0.41 (0.0753) 1.13 (0.1034)Lived with BothParents at 14
0.24 (0.1531) 0.60 (0.2296)
Mother HS Grad 1.76 (0.1319) 2.64 (0.2950)Mother CollegeGrad
2.52 (0.4200) 5.45 (0.4820)
Estimated standard errors are in parentheses. Numbers which appear without standard errors arederived from other estimated parameters (with standard errors) or from the simulated data. Allparameter estimates were estimated within the maximization of the likelihood function except for theintitial conditions for parental background which were estimated seperately outside the model in orderto start the simulation procedure with simulated agents who have similar family backgroundcharacteristics to the sample observed in the NLSY data (the logits for these characteristics aredesigned to produce the same proportion of agents with each background characteristic and thecorrelations between those characterisitics).